#IlyaSutskever discusses the challenges of #AI #modelgeneralisation, comparing it to #humanlearning. He suggests that the current focus on #RLtraining, driven by evaluation metrics, might be limiting model adaptability. Sutskever proposes that expanding training environments or improving generalisation from pre-training data could enhance model performance across diverse tasks. https://www.dwarkesh.com/p/ilya-sutskever-2?eicker.news #tech #media #news